[MfA]

Originally Posted by CarstenS

People have been told time and again (not by AMD directly though!!) that thanks to VLIW5/4 you get way higher compute density than with other approaches.

The main advantage is the potentially shorter runtime of tasks and thus less use of temporary storage. Maybe superscalar can be done just as cheap, although NVIDIA didn't seem to manage, but going pure scalar (from a work item perspective) has a negative impact on the use of temporary storage (needs more work items running concurrently).

Ideally you would be able to do both ... VLIW or SIMD, depending on whatever code would be running at the time.

[rpg.314]

http://semiaccurate.com/2012/06/12/a...-and-mediatek/
Are there any more technical details?

[lanek]

Im surprised there's not so much article about it yet... HSA seems really an important thing for future of computing and system.

[CarstenS]

Maybe because AFDS is still going on and people are busy partying... errr, attending sessions and meetings.

[rpg.314]

Intel, nvidia and qualcomm's absence is a downer.

Arm and Amd means intel wouldn't be too happy joining hsa foundation. But ARM's presence would suggest nvidia and qualcomm would want to join.

[lanek]

Originally Posted by CarstenS

Maybe because AFDS is still going on and people are busy partying... errr, attending sessions and meetings.

yes i can understand they need a good night of recuperation before write something..

[Alexko]

Originally Posted by rpg.314

Intel, nvidia and qualcomm's absence is a downer.

Arm and Amd means intel wouldn't be too happy joining hsa foundation. But ARM's presence would suggest nvidia and qualcomm would want to join.

AMD's presence probably explains NVIDIA's reluctance too.

But I'm not sure what's keeping Qualcomm from joining, or even Samsung, for that matter.

[rpg.314]

I just noticed that extension of coherency to discrete GPU has been silently dropped from the roadmaps.

[AlexV]

Originally Posted by rpg.314

I just noticed that extension of coherency to discrete GPU has been silently dropped from the roadmaps.

They were talking about it as if it was still under consideration, although I can't say there was that much meat on the HSA bones this year...or maybe I always ended up in the wrong place at the wrong time

[rpg.314]

Well, seeing it gone on the blessed slides can't be good...

[AlexV]

Originally Posted by rpg.314

Well, seeing it gone on the blessed slides can't be good...

The common meme at AFDS was about ninjas being evil and the common developer being the one you should care about...so maybe somebody got scared about coherency being too ninja-like. At any rate, even if it had been on the slides it would have been risky to bet on things so far out, if prior experience is anything to go by, IMHO.

[rpg.314]

Originally Posted by AlexV

The common meme at AFDS was about ninjas being evil and the common developer being the one you should care about...so maybe somebody got scared about coherency being too ninja-like. At any rate, even if it had been on the slides it would have been risky to bet on things so far out, if prior experience is anything to go by, IMHO.

Well, on the scale of system in question, I would have thought that common developers would like coherency while explicit DMA's would be for ninja's.